We endow Large Language Models (LLMs) with fine-grained self-evaluation torefine multi-step reasoning inference. We propose an effective promptingapproach that integrates self-evaluation guidance through stochastic beamsearch. Our approach explores the reasoning search space using awell-calibrated automatic criterion. This enables an efficient search toproduce higher-quality final predictions. With the self-evaluation guidedstochastic beam search, we also balance the quality-diversity trade-off in thegeneration of reasoning chains. This allows our approach to adapt well withmajority voting and surpass the corresponding Codex-backboned baselines by$6.34\%$, $9.56\%$, and $5.46\%$ on the GSM8K, AQuA, and StrategyQA benchmarks,respectively, in few-shot accuracy. Analysis of our decompositional reasoningfinds it pinpoints logic failures and leads to higher consistency androbustness. Our code is publicly available athttps://github.com/YuxiXie/SelfEval-Guided-Decoding.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)